Skip to main content

Stable Diffusion

Generate and control brand-grade images with a latent diffusion model — prompts, ControlNet, LoRA, inpainting, upscaling — with the trade-offs that come with them.

Course Duration: 6h​

What You'll Learn​

  • Explain the three pieces of a latent diffusion model — autoencoder, U-Net, scheduler — and why SDXL scales them up
  • Write positive and negative prompts that actually change the image, and know what a negative prompt cannot fix
  • Pick a step count, a guidance scale and a random seed with reproducibility in mind
  • Chain text-to-image and image-to-image, tune denoising strength for controlled variations
  • Retouch a detail with a mask (inpainting) and extend a canvas for a banner format (outpainting)
  • Steer composition with ControlNet using pose, edges and depth preprocessors
  • Teach the model a house style with a LoRA of 20 to 40 images, and spot overfitting early
  • Upscale for print, restore faces, cap memory and read a compute-cost budget
  • Respect model and LoRA licenses, watermark provenance and stay inside acceptable-use limits

Prerequisites​

  • Course 15 — Generative AI
  • Working Python 3.10+, a GPU with at least 8 GB of VRAM (12 GB recommended for SDXL), and the diffusers library installed

Course Modules​

  1. Latent diffusion: autoencoder, U-Net, scheduler
  2. Positive and negative prompts
  3. Steps, guidance scale, random seed
  4. Text-to-image and image-to-image
  5. Inpainting and outpainting
  6. ControlNet: pose, edges, depth
  7. LoRA and textual inversion for a style
  8. Upscaling and face restoration
  9. Compute cost and memory optimization
  10. Rights, provenance and acceptable use

The running project​

Across the ten modules you produce the visuals for a small furniture brand. A reference chair is declined into ambience shots (text-to-image), color variants (image-to-image), a retouched armrest detail (inpainting), an extended banner frame (outpainting), a controlled pose in a staged photo (ControlNet), a learned house style (LoRA), and finally a print-ready upscale. The tooling is the diffusers Python library end to end, with the equivalent moves in a graphical UI (ComfyUI) noted in a callout. The reference model is SDXL, with SD 1.5 mentioned when a machine is modest.

Assessment and certificate​

The course ends with a 40-question exam covering every module. On success, a certificate of completion is issued; its number is verifiable on the platform.

Free courses, by contrast, end with a 5-question quiz and a preview of the certificate, without certification.